docs(plans): codebase audit 2026-09-02 with ranked remediation plan - #685
Closed
Kaden-Schutt wants to merge 3 commits into
Closed
docs(plans): codebase audit 2026-09-02 with ranked remediation plan#685Kaden-Schutt wants to merge 3 commits into
Kaden-Schutt wants to merge 3 commits into
Conversation
Twelve read-only audit slices over master 8cd15a6 (loader/daemon, generate, runtime, config/topology, arch crates x3, dispatch/kernels, quantize, redline/spec, cli/tui/scripts, docs/issues), synthesized into one report. Three failure classes dominate: the loader admits combinations the engine cannot run (six #683-class cases), reset/rollback is canonical on three routes and hand-rolled or absent elsewhere, and docs/comments describe the design rather than the code. Five findings re-derived by reading the cited lines; the rest carry the slices' citations.
Five read-only slices over the 34 qt=44 translation units plus the Rust encoder/dispatch/launchers/parity, and an independent pass over every half-select site. Verdict: the format is sound end to end. One real defect (llama::is_batchable_la admits V2 WMMA prefill on gfx12 only while qwen35 admits gfx11+gfx12; both claim lockstep), one test gap (mq4v2_gemm_parity's Gaussian fixture cannot discriminate a wrong half-select), spec section 9 stale on MoE/XBATCH/K-step.
Five read-only slices over the draft runtime, generate loop, verify/replay, attention_dflash kernels, and pairing/discovery. The machinery is sound; the problems are around it: no draft auto-pairing exists (dflash_mode auto/on without an explicit path is plain AR; AGENTS.md and MODELS.md describe a matcher that is not in the tree), draft VRAM is invisible to admission, draft ctors leak on partial failure, make_spec_emitter Err skips rollback, and the entry/loop ctx-cap predicates disagree by block_size.
This was referenced Sep 3, 2026
Open
Collaborator
Author
|
Superseded: the ranked remediation rows were executed on beta as individual PRs (#687 admission, #688 free_gpu + init_ep, #686/#691 DFlash, 103a5f1 docs drift). Still open from the plan — GL quant 38/39 dense refusal, DeviceBuffer Drop, MQ8 GEMV rotation — belong in issues, not a plan doc pinned to 8cd15a6. Ships via #730. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Point-in-time audit of master
8cd15a62b: twelve read-only slices, one synthesized report (docs/plans/audit-2026-09-02/README.md) with a ranked remediation plan, plus the per-slice reports. Every finding is cited topath:lineat that commit.Headline: the failures cluster into three classes — the loader admits (model, topology, quant) combinations the engine cannot run (six cases of the #683 species), reset/rollback is canonical on AR/dense-TP/DS4 and hand-rolled or absent on EP error exits, vision cancel, and the bring-up archs, and docs/comments describe the design rather than the code. Also verified: an MQ8 GEMV family path that panics,
LlamaWeights::free_gpuleaking PARO/AWQ sidecars on every reload, and EP loads recording aTpmesh.Which surface(s) does this touch?
docs/plans/audit-2026-09-02/*, one row indocs/INDEX.mdNo code. No hardware route.